We publish every anti-cheat signal Ishavi collects during a live interview, the threshold that triggers it, and what we do with it. A single flag never auto-rejects a candidate -- multi-signal coincidence plus a human reviewer is the only path to escalation.
Tier 1 (on-device camera analysis) ships post-MVP. Today we use behavioural-only signals.
These are the Tier 0 detectors active in the current build. Anything not on this list is not collected. We add new detectors only after listing them here first.
| Signal | What it observes | Threshold | Action taken |
|---|---|---|---|
| Tab hidden | Browser tab was hidden (you switched tabs or minimised the window). | Any occurrence; the cumulative duration is reported. | Logged. Multi-signal trigger only at >=60s cumulative hidden time. |
| Window blur | Keyboard focus left the browser without a tab switch. | Any occurrence. | Logged. Counts toward multi-signal coincidence, never standalone. |
| Paste | Clipboard paste into any focused element on the interview page. | Any occurrence; the target element is recorded. | Logged + visible on the recruiter scorecard timeline. |
| Fullscreen exit | You entered fullscreen mode and then exited (only fires after you opted into fullscreen). | Any occurrence. | Logged. |
| Developer tools open | Heuristic: difference between window outer/inner dimensions exceeds 160 px (consistent with the devtools pane open). | Refires at most once per minute to avoid noise. | Logged. |
| Multi-display detected | Browser reports `screen.isExtended = true` (a second monitor is attached). | Once per session at the start. | Logged only. Owning a second monitor is not a failure mode. |
| Face absent | On-device face detector reports no face for an extended interval. Runs only on supported browsers; disabled in the EU. | Confidence < threshold for >= 3 seconds. | Logged. Reviewable clip attached. |
| Multiple faces | On-device face detector reports more than one face. Disabled in the EU. | >= 2 faces detected for >= 3 seconds. Static-poster detection runs first. | Logged. Reviewable clip attached. |
| Audio input change | Active microphone changed mid-interview. | Any occurrence. | Logged. |
| Virtual camera suspected | Combined detector: virtual-camera driver name, frame-latency variance, and landmark-jitter analysis. Disabled in the EU. | All three sub-signals agree. | Logged + flagged for recruiter review. |
| Knowledge depth failed | An unscripted follow-up to your own answer was not answered or was answered with a contradiction. | Server-side LLM judge marks the follow-up as failed. | Logged + folded into the rubric score (not standalone). |
| Question-bank leakage | The same question text appeared verbatim for two candidates AND the second candidate responded unusually fast. | SHA-256 hash match + response-time outlier. | Logged + recruiter notification. |
| Lag-loop outlier | Pause between the AI finishing a question and your first speech, measured against the population baseline for that question's difficulty. | >= 3 seconds across >= 3 consecutive questions (population baseline, not your own). | Logged. |
| Identity embedding drift | On-device face embedding drifted from the enrolment selfie. Disabled in the EU. | Rolling median distance > 0.6 over a 30-second window. | Logged + flagged for recruiter review. |
Any one of the signals above can fire for an innocent reason -- a notification interruption, a sibling walking past the camera, a flaky microphone. The platform's decision-making code requires multi-signal coincidence (two or more signals firing in the same window) anda final score that falls below the rubric's minimum before flagging an interview for human review. The final pass/no-pass decision is always taken by a human; the AI's recommendation is advisory.
European Union / United Kingdom
EU AI Act Art. 5(1)(f) prohibits emotion inference and biometric categorisation in workplace contexts. The face-detection, multi-face, identity-drift, and virtual-camera detectors are forcibly disabledwhen the tenant region is EU or UK or when the candidate's jurisdiction is EU. The behavioural detectors (tab hidden, window blur, paste, lag-loop, devtools open) keep running -- they are not emotion or biometric signals.
United States
All Tier 0 signals are available. NYC Local Law 144 candidates see an explicit notice at least ten business days before the AEDT is used in their hiring. Illinois AIVIA candidates see the reviewer-instrumentation disclosure on the scorecard. Colorado candidates are not currently accepted while the platform completes its CO AI Act compliance review.